Audience
← Back to Course Catalog /10 · Tier III

TDA for Robust & OOD-Aware ML

For ML engineers and AI-safety researchers. Adversarial robustness, OOD detection via topological signatures, certified robustness via topological structure, topology-aware training. Launched once AIRINA's published work in this area gives the standing to teach it.

Adversarial robustness and out-of-distribution detection both have a topological flavor: small perturbations move points off the data manifold, OOD examples sit in different connected components of the support, distribution shift changes the global shape of the data. Several recent papers exploit this. Several others overclaim it. This week is for working ML practitioners who want a sober view of what topology actually contributes.

This program opens once AIRINA's own published work in the area is mature. The honest framing: we are not running it yet because we do not yet have the standing to teach it. Express interest now and we will reach out when the cohort opens.

Program Overview

Five consecutive days, in-person at AIRINA Labs in Cotonou, bilingual EN/FR. Taught by AIRINA researchers in TDA and applied ML, plus an invited robustness researcher from the international AI-safety community (named four weeks before each cohort). The program opens for its first cohort once AIRINA's own publications in the area are in place.

Program structure

  • Days 1–4 — lectures, labs, case studies. The current landscape of adversarial robustness, topological signatures of OOD, certified robustness, and topology in the training loop.
  • Day 5 — final project + defense. Participants pick a robustness or OOD problem from a curated list (or bring their own with prior approval), build a topology-enhanced detector or robust trainer, and defend it against an honest baseline.
  • Cohort size. ~12. Application-based.
  • Materials. Reading list and pre-course exercises 4 weeks ahead.

Certificate

Successful completion of the final project earns an AIRINA Robust ML with TDA certificate — graded final. The grade is on whether your comparison would survive a paper review: a topological method on a real robustness or OOD-detection problem, with a strong classical baseline in the same evaluation harness.

Learning Outcomes

By the end of the program, participants will be able to:

  • Compute topological signatures of activations and inputs, and use them as features in a robustness pipeline.
  • Build a persistence-based OOD detector and benchmark it honestly against deep ensembles, energy-based detectors, and Mahalanobis-distance baselines.
  • Apply differentiable persistence in a robust-training loop, with awareness of the subgradient issues.
  • Read a recent "topology gives robustness" paper and identify whether the claim survives proper baselines.
  • Decide when topological methods are warranted for a robustness problem and when they are not. Both happen.

Program curriculum

Day 1 · Adversarial robustness — current landscape

Threat models, attacks (PGD, AutoAttack, transfer attacks), and defenses (adversarial training, randomized smoothing, certified bounds). Where each method works and where it does not. The honest picture of the robustness-accuracy trade-off.

Day 2 · Topological signatures of OOD examples

Persistence diagrams of activations on in-distribution vs OOD inputs. Building a persistence-based OOD detector. Benchmarking against deep ensembles, energy scores, and Mahalanobis on CIFAR-style and tabular benchmarks.

Day 3 · Certified robustness via topological structure

Where geometric/topological properties of the data give certificate-level guarantees, and where they merely give empirical robustness. Limitations and open problems.

Day 4 · Topology in the training loop

Differentiable persistence applied to robust training; topological regularizers; the cost in clean accuracy and how to think about it.

Day 5 · Final project + defense

Participants pick a robustness or OOD problem from a curated list (or bring their own with prior approval), build a topology-enhanced detector or robust trainer, and defend it against an honest baseline in a 30-minute session.

Who Should Attend

This program is for ML practitioners who want to know what topology actually contributes to robustness and OOD-detection problems they ship.

  • ML engineers and research engineers working on robustness, OOD detection, model monitoring.
  • AI-safety researchers interested in topological methods.
  • Risk and security teams at financial institutions and beyond who deploy ML in adversarial settings.

Prerequisites

  • ML practice. Strong ML practitioner background: PyTorch, training loops, evaluation discipline.
  • Robustness toolkit. Working familiarity with adversarial attacks (PGD, FGSM, AutoAttack) and OOD-detection baselines.
  • TDA. Some prior TDA exposure — Training /01 is a recommended prerequisite or equivalent.

Selection

Application-based. The program opens once AIRINA's research output in the area is mature; register your interest and we will reach out when a cohort is scheduled.

Brochure

The detailed program brochure (PDF, EN/FR) is sent on request — including day-by-day curriculum, faculty profiles, the curated project list, and the cohort calendar.

To receive the current brochure, write to contact@airina.africa with "Robust ML — brochure request" in the subject. The brochure is updated each cohort; we send the version current at the time of your request.